acceptodds
Under review as a conference paper at ICLR 2027

Epigenetic Age Prediction with Sequence-Conditioned Co-Methylation Graph Networks

Abstract

Epigenetic clocks estimate biological age from DNA methylation profiles and are widely used in ageing research. Graph-based clocks model relationships among CpG sites but typically omit the local DNA sequence surrounding each site, despite its potential influence on methylation. Using sequence context requires careful interpretation because the sequence of each CpG site is fixed across individuals: improved prediction may reflect informative local context or merely provide a way to distinguish CpG sites. We introduce a sequence-conditioned graph clock that jointly models sample-specific methylation, local sequence context, and relationships among CpG sites, with sequence conditioning accounting for only of the model's parameters. To determine what drives any improvement, we use matched comparisons that distinguish the contributions of sequence context and CpG relationships while testing whether gains depend on the correct sequence–site correspondence. Across 37 blood methylation studies, sequence conditioning consistently improves biological-age prediction, reducing mean test MAE from to years. This improvement largely disappears when sequences are assigned to the wrong CpG sites, showing that correct local context, rather than CpG-site identity alone, accounts for most of the gain. Prediction averaging further reduces MAE to years, and the model performs favourably in an exploratory application to ovarian cancer samples. These findings show that local DNA sequence provides useful information for co-methylation graph clocks and highlight the importance of distinguishing sequence information from CpG-site identity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
DNA methylation, epigenetic aging, epigenetic clocks, graph neural networks, co-methylation graphs, genomic sequence, sequence conditioning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LkUqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper developed a graph clock conditioning on DNA sequences to improve predicting biological age. The proposed approach enables to jointly model multiple molecular features related to DNA methylation, including individual CpG sites, sequence-specific context, and CpG-CpG relationships. Through matched comparisons, the paper singled out the feature-specific contributions from sequence-specific context and CpG-CpG interactive relationship, while assessing extra gains by including local sequence context in the model.

Critical strengths

This paper demonstrates the added benefit for biological age prediction in the framework of graphic models from the use of embedding information extracted from local sequence-specific context and CpG relationships, along with the use of neural networks in prediction model formulation. The factorial design to evaluate graph contribution is well reasoned as this design allows to isolate the effect of graph-message-passing from the effect of DNA sequence construct. In this way, this paper attempted to provide quantitative evidence on the influence of both sources of information in the building of the prediction model.

Critical weaknesses
  1. The proposed model shows little improvements over simpler methods such as the Elastic Net Regularized Linear Model (ENReLM), despite its much greater complexity due to the invocation of overparametrization and nonlinearity. It also has shown higher run-to-run variability, which raises concerns about stability. Ensembling helps add extra information to potentially gain predictive power, but it is unclear whether these gains justify the additional costs of obtaining DNA sequence data and the required neural network computing in comparison to the classical ENReLM. Overall, the results reported in this paper do not convincingly show that the proposed method offers a meaningful advantage over existing epigenetic clocks.
  2. The proposed architecture combines sequence encoding, graph message passing, and an eight-layer MLP. It is unfortunate that the paper does not sufficiently justify, biologically, the need for such a complex combo of signals in the biological age prediction. The factorial ablations examine the contributions of sequence conditioning and graph propagation, but leave other architectural choices largely unexplored. The proposed approach is also harder to interpret biologically than simpler models, especially the ENReLM, which may limit its practical usefulness as seen by practitioners.
Questions for authors
  1. What’s the biological rationale behind the use of the DNA sequence for predicting biological age, when methylation data are already reported in the literature as primary molecular aging features? Aging, especially aging heterogeneity is heavily influenced by exposures to environmental and contextual factors, and there lacks clear biological basis that links the DNA sequence with such exposures.
  2. Why neural networks, particularly the MLP component in terms of overparametrization and nonlinearity, should improve your performance? ENReLM is known to be highly adaptive to local data, and the resulting sparse linear prediction model is also known to be robust and generalizable such as Horvath’s clock. It seems that the proposed neural network model does not really help for the prediction power improvement.
  3. There are hundreds of biological age clocks available in the literature, what’s the reason for picking the two benchmark clocks in the comparison as shown in Table 1? The presented comparison to the two chosen benchmark clocks is not really fair, because typically a transfer learning step is used to calibrate an existing clock to make it more adaptive to the local target data. This step is not considered in this paper, resulting in an unfair comparison.
7 comments

We thank the reviewer for the detailed, practice-oriented review and for recognizing the factorial design.

Weakness 1 — Improvement over Elastic Net, variability and cost

Reviewer concern. The gain over Elastic Net is small relative to the added complexity, runs are variable, and the cost may not be justified.

We added an Elastic Net with Horvath’s age transformation, the strongest linear model we tested (MAE 3.360 years on the original split). A single run of our model is on par with it (+0.05 years, not significant), whereas our three-seed ensemble improves on it by 0.23 years [0.07, 0.40] in all five folds, both for samples aged 20 or younger (0.68 vs 0.88) and for adults (4.15 vs 4.40). Ensembling also reduces the run-to-run standard deviation from 0.21 to 0.085 years. The sequence adds no data cost, because each probe’s 122-bp context is listed in the public Illumina manifest, and no inference cost, because the sequence features are fixed per site after training.

Weakness 2 — Architectural complexity and interpretability

Reviewer concern. The combination of sequence encoding, message passing and an eight-layer readout is not justified, and the model is hard to interpret.

As the reviewer notes, the model has three components, and the factorial ablations already isolate the contributions of two of them, sequence conditioning and graph propagation. Each component has a distinct role:

  • Sequence encoder. A site's local sequence shapes how its methylation is established and maintained; for example, GC skew protects CpG-island promoters from methylation. The encoder turns this context into a prior for each site, and its contribution is confirmed by the factorial ablations and the sequence-reassignment control.
  • Graph message passing. Age-related methylation changes occur in co-methylated modules, and single-CpG measurements are noisy, so each site borrows evidence from its co-methylated neighbours. Its contribution is isolated by the factorial ablations.
  • MLP readout. This is a standard component rather than a design choice of ours. MLP readouts are the usual final stage of graph neural networks such as GIN and PNA and of graph-based clocks such as GraphAge, and deep clocks such as AltumAge and DeepMAge are themselves MLPs. Like the coefficient vector of a linear clock, the readout maps the site representations to a single age, but non-linearly. It is identical in every arm of the factorial design and on its own forms the design's baseline, so the gains from sequence and graph are measured on top of it. Beyond its input layer, which reads all 20,318 sites, the whole network has fewer than 0.9 M parameters.

Interpretability. A sparse linear clock offers a list of sites and weights, but this list is unstable: Horvath's and Hannum's clocks share only six CpGs, and in our data the Elastic Net selects 1,731 to 3,147 sites per fold, only 776 of them in all five folds, although the training partitions overlap by three quarters. Our model can be interpreted at two levels. The sequence pathway shows that the clock relies on GC content and GC skew, which mark unmethylated CpG islands, and the contribution of each site to each prediction can be computed with standard attribution methods and summed over co-methylated neighbourhoods.

Question 1 — Biological rationale for DNA sequence

Reviewer concern. What is the biological rationale for using DNA sequence to predict biological age, when methylation is already reported as the primary molecular feature of ageing? Ageing, especially its heterogeneity, is heavily influenced by environmental and contextual exposures, and there is no clear biological basis linking DNA sequence to such exposures.

Sequence does not replace methylation as the ageing feature; it tells the model how to read each site's methylation. A site's local sequence determines its typical methylation level and how it tends to change: GC-skewed CpG-island promoters are protected from methylation, whereas CpGs flanked by A or T lose methylation with age. Our trained models learn this kind of information: their shared sequence signal tracks GC content and GC skew and reflects each site's typical methylation level.

This is useful even though sequence does not reflect exposures. In a methylation clock, exposures are visible only through the methylation changes they cause, and how a site responds depends on its sequence context: age- and exposure-related changes differ systematically between CpG islands and other regions. Sequence therefore provides the reference for reading these changes. The same measured value indicates a gain of methylation at a site whose sequence keeps it unmethylated, but a loss at a site whose sequence favours methylation.

Question 2 — Why neural networks?

Reviewer concern. Why should an over-parameterised, non-linear model outperform a sparse linear clock?

Methylation changes non-linearly with age, rapidly in childhood and slowly in adulthood, which is why Horvath’s clock needs a logarithmic age transformation below age 20. A neural network learns such non-linear relations, and interactions among sites, directly from the data instead of through a fixed transformation, and over-parameterisation is controlled by dropout, weight decay and validation-based checkpoint selection. Our results support this: on the original split, the neural models outperform the strongest linear model we tested, an Elastic Net with Horvath’s age transformation, in all five folds (tuned MLP by 0.17 years, DeepMAge by 0.13 and our three-seed ensemble by 0.23). Our ensemble improves on it both for samples aged 20 or younger (0.68 vs 0.88) and for adults (4.15 vs 4.40).

Question 3 — Choice of clocks and transfer learning

Reviewer concern. Why Horvath and AltumAge, and why without local calibration?

Table 1 is not limited to two clocks: it compares our model with six models. Four are trained on our partitions, an Elastic Net, a sequence-agnostic MLP, ResnetAge and a sequence-agnostic PNA, and two are published clocks used as external references, Horvath and AltumAge. Together they cover the main families of chronological-age clocks: penalised linear (Elastic Net, Horvath), deep feed-forward (MLP, AltumAge), convolutional (ResnetAge) and graph-based (sequence-agnostic PNA). We chose models that predict chronological age, as ours does, and that represent distinct model families. Horvath’s clock is the most widely used linear clock, and AltumAge is a deep clock defined on exactly our 20,318 sites. Clocks with other targets, such as PhenoAge, GrimAge and DunedinPACE, are not like-for-like comparisons.

Following the reviewer’s suggestion to adapt the clocks to the local data, and to ensure a fair comparison, we retrained both published clocks on our data while keeping their original design, so that every model in the comparison is trained on our local training data. Both use the same training partitions, validation-based model selection and test set as our model, and retraining also removes any overlap between the published clocks’ training data and our test samples.

MAE (years), original splitRetrained from scratch
Horvath’s clock3.949
AltumAge3.588
Ours3.405

The rebuttal addresses some of the concerns in the review reports, with a few major points remaining unresolved. Regarding the performance comparison against Elastic Net, the proposed approach demonstrates an advantage only when a three-seed ensemble model is fitted. Particularly, an improvement of 0.23 years (2-3 months) is not biologically significant. This improvement comes at a substantially higher cost, both in terms of the additional data requirements associated with sequential inputs and the computational resources needed to train and fit a complex neural network architecture. Regarding the responses to the interpretability concerns, the author provided a detailed discussion of the rationale behind the design choices for each module, as well as their discussion of the possibility of interpreting model predictions using standard attribution methods. The concern remains: the paper did not include a concrete example of applying attribution methods for prediction interpretability demonstration. Regarding the modeling of nonlinear relationships using neural networks, there lacks comparison against other classes of nonlinear models with simpler structures, such as Random Forests, XGBoost, or generalized additive models, to further support the benefits of choosing the neural network approach. Despite these clarifications above, the manuscript itself did not incorporate any changes. Given these remaining concerns, the original rating is adequate.

Reviewer_gj5qOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper deals with learning epigenetic clocks, which are models that estimate the biological age of samples from DNA methylation. It proposes a way to add the DNA sequence around each CpG site (potential methylation site) as an additional covariate on top of the graph already used locality of the sites. The paper shows that this approach leads to improvements in age estimation, and improve age estimation by ~0.1 years (this is a statistically significant effect when conditioning on training-splits and seeds, but is comparable to the noise of individual runs). The authors demonstrate that the local context have a direct effect (rather than simply id'ing the cpg sites) by permuting the context labels and showing that performance decays back.

Critical strengths
  • In the context of methylation-based biological clocks, the paper developes a method to combine the effects of the local sequence around the cpg site into the model to graph based biological clock models (the "PNAs"). They shows (using a nice control) that this addition improves the model. This is a contribution to current modeling of methylation-based biological clocks.
Critical weaknesses
  • It is not clear in what sense is it important that there is some additional signal in prediction accuracy of biological age when local sequence is added into the model. Are the benefits in model accuracy important in that they allow something new or an increase in predicitive capacity that unlocks new usage? Or alternatively is this surprsing biologicially or allows us to learn something about methylation and biological clocks? I have not been convinced of either possibilities.

  • I am not convinced that there is substantial methodolgical novelty. Both the sequence model and the PNA have been proposed previously in these contexts. This is a combination of both approaches, that achieves modest gains. So I am not sure there is much to learn beyond the technical finding. Also, there is no discussion of what sequence information has been uncovered, so no new biological hypothesis to look at.

Questions for authors

Questions: Is the approach of adding the sequence context new or surprising? What are key insights in using this (Figure 1) specific architecture to add in context? Can you articulate in what way this method advances our understanding or use of methylation-based molecular clock .

Suggestions for authors: The constants in eq (1) seems arbitrary, and the equation itself is hard to read.

4 comments

We thank the reviewer for the clear summary, for highlighting the permutation control and for judging the paper sound. We address each concern below.

Weakness 1 / Question 3 — Practical and biological significance

Reviewer concern. Does the extra accuracy unlock a new use, or reveal something about methylation clocks?

Yes, in two respects. First, biological-age accuracy matters for age acceleration. Clocks systematically under-predict older individuals, so age acceleration partly reflects age itself and can be mistaken for a disease effect. In our ovarian cancer application (Section 5.6), our model has the smallest under-prediction in the three oldest groups (−10.3 years above 80, vs −13.7 for AltumAge and −15.5 for Horvath), so disease-related deviations are less confounded by calibration bias. This is a prerequisite for using age acceleration in disease studies and, potentially, for early detection. Second, the gate is one sequence-determined weight per site, so the clock's use of sequence can be inspected directly.

Weakness 2 — Methodological novelty and What sequence information is learned

Reviewer concern. 1. Both the sequence model and PNA have been proposed before, and their combination yields modest gains, so there may be little to learn beyond the technical finding. 2. The paper does not discuss what sequence information has been uncovered, so it offers no new biological hypothesis.

For the first concern, our work is application-driven: its aim is to advance biological-age prediction. Its contribution lies in identifying information that existing clocks leave unused, bringing appropriate tools to exploit it, and verifying that it is exploited as intended.

  1. A new source of information that improves accuracy, age-acceleration estimates and interpretability. Most clocks characterise each CpG site only by its measured methylation, and graph-based clocks add correlations between sites and coarse annotations, but neither uses the sequence surrounding each site, although this sequence shapes how methylation is established and maintained. To our knowledge, ours is the first clock that jointly models sample-specific methylation, the local sequence of each CpG site and explicit co-methylation relationships, and doing so brings three benefits. First, higher accuracy: our model has the lowest MAE in Table 1 and remains more accurate than Horvath’s clock and AltumAge retrained from scratch on our data (three-seed ensembles: 3.128 vs 3.948 and 3.394 years, lower in all five folds). Second, more reliable age acceleration: in the ovarian cancer application, our model has the smallest under-prediction in the three oldest age groups (−10.3 years above age 80, vs −13.7 for AltumAge and −15.5 for Horvath’s clock), so disease-related deviations are less masked by age-related bias. Third, interpretability: each site’s sequence features can be analysed directly, which showed that sequence provides a prior on each site’s methylation level, linked to GC content and GC skew, an insight unavailable to clocks based on methylation alone.
  2. A design matched to the structure of methylation data. Our architecture respects the asymmetry between methylation and sequence: methylation remains the only sample-specific input, and sequence, which is identical in every sample, only conditions how each site is represented and propagated on the co-methylation graph. A lightweight convolutional encoder, trained jointly with the clock and without pretraining, suffices for this purpose and adds only 0.05% of the parameters.

For the second concern, we analysed the sequence pathway of all 15 trained models (five folds × three seeds). What the models learn in common is the GC content and GC skew (the excess of G over C) of each site's 122-bp context: these two features explain 52% of the sequence signal shared across models, and 4-mer composition explains 66%, compared with 26% for GC content alone. GC skew is a known mark of unmethylated CpG-island promoters, where it promotes R-loop formation that protects DNA from methylation. Consistent with this mechanism, the learned signal is correlated with a site's typical methylation level (|ρ| = 0.19) but not with its correlation with age (|ρ| ≤ 0.03). Sequence therefore acts as a prior on each site's baseline methylation level. In the trained models this prior is carried by the two-dimensional sequence projection, while the gate remains close to 1, so methylation values enter message passing essentially unscaled.

Question 1 — Is adding sequence context new or surprising?

Reviewer concern. Is the approach of adding sequence context new or surprising?

It is new for epigenetic clocks: to our knowledge, no previous clock jointly models sample-specific methylation, the local sequence of each CpG site and explicit co-methylation relationships. Two findings were less expected. First, a lightweight encoder trained only on the age-labelled methylation profiles, without the large-scale pretraining used by sequence-aware foundation models, is enough to extract useful sequence information. Second, what it extracts is not an age-specific motif but a prior on each site's methylation level, tied to GC content and GC skew.

Question 2 — Key insight of the architecture

Reviewer concern. What is the key insight of the design in Figure 1?

The design separates what varies between individuals from what is fixed for each site. Methylation is the only sample-specific input, whereas the sequence of a site never changes, so sequence is used only to condition how each site is represented. A shared CNN maps each site's 122-bp context to a gate in (0, 1) that multiplies the methylation value before message passing, so sequence determines how strongly each site's signal enters the messages exchanged on the co-methylation graph. Because the gate is a shared function of sequence rather than a free parameter per site, sites with similar contexts receive similar gates, which biases the pathway towards sequence content rather than site identity; the reassignment control tests whether this holds. The whole sequence pathway adds only 0.05% of the parameters, and the scalar gate keeps its effect readable as one number per site.

Suggestion — Eq. (1)

Reviewer concern. The constants in Eq. (1) seem arbitrary, and the equation is hard to read.

We thank the reviewer and will rewrite Eq. (1) with an explanation of its constants. The equation defines the edges of the co-methylation graph, which is built separately for each fold from the training samples only. Nodes are the 20,318 CpG sites, and two sites are connected when their methylation levels are strongly correlated across training individuals, with a slightly lower threshold for sites that are close in the genome. In the revision we will write the rule as a single threshold that depends on genomic proximity: sites i and j are connected when |rᵢⱼ| > τᵢⱼ, where rᵢⱼ is their Pearson correlation over the training samples and

  • τᵢⱼ = 0.70 for sites on different chromosomes,
  • τᵢⱼ = 0.68 for sites on the same chromosome more than 100 kb apart,
  • τᵢⱼ = 0.66 for sites on the same chromosome within 100 kb.

Each edge carries rᵢⱼ and two indicators, one for a shared chromosome and one for a shared gene.

The constants encode a simple prior rather than tuned hyperparameters. The base threshold of 0.70 retains only strong co-methylation and yields a sparse graph, with 2.1 to 2.4 million undirected edges among the 20,318 sites in each fold. Co-methylation is strongest between nearby CpG sites and weakens with genomic distance (Affinito et al., 2020), so a correlation just below 0.70 is more likely to reflect a shared regulatory context when the two sites lie on the same chromosome, and more so when they are within 100 kb. The relaxations are therefore graded by proximity and kept small (0.02 and 0.04). They add only 1.2% of the edges (fold 1), so the graph is essentially determined by the base threshold.

Reviewer_gj5qOpenReview ↗

Thank you, maintaining score

Thank you for the clarifications.

I feel that as a practical application, the benefits in accuracy are very modest. From a general scientific interest, I don't think the insights or novelty is sufficient for this conference.

I therefore maintain my score.

We thank the reviewer. To make it clear, on accuracy, the gain is consistent: with three-seed ensembles, our model outperforms the age-transformed Elastic Net, random forests, XGBoost and GAMs in all five folds. On novelty, beyond clocks, the paper shows how a reassignment control separates the content of a fixed node input from its role as a node identifier, and it finds that sequence acts as a prior on each site's baseline methylation.

Reviewer_M9DcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper combines local DNA sequence context with co-methylation graph networks to improve epigenetic age prediction and investigates the contributions of sequence information, graph propagation, and ensembling.

Critical strengths

Matched ablations, sequence-reassignment controls, and repeated training provide useful evidence that correctly paired sequence context contributes beyond CpG-site identity.

Critical weaknesses

Modest graph-specific gains, limited comparisons with strong external methods, and the absence of unseen-study evaluation leave the architecture’s practical advantages insufficiently established.

Questions for authors

oes the proposed model outperform strong sequence-aware non-graph models and existing graph-based clocks under matched computational budgets and held-out-study evaluation?

1 comment

We thank the reviewer for recognizing the matched ablations, the sequence-reassignment control and the repeated training.

Weakness / Question — Held-out studies, strong baselines and matched compute

Reviewer concern. Does the model outperform strong sequence-aware non-graph models and existing graph-based clocks under matched compute and held-out-study evaluation?

Yes, our model has achieved the best performance compared with other models. To answer this question, we added leave-studies-out validation: the 37 studies have been split into 5 groups. Specifically, four of them will be used as training data, and one of them is used as testing data. Here are the detailed results.

MAE (years), unseen studiesMAE↓
Ours4.35
Horvath + linear calibration5.24
Elastic Net + age transform5.72
GraphAge (sequence-agnostic PNA)5.74
DeepMAge5.91
Horvath, refit on its own CpGs6.12
Horvath (published)6.16
AltumAge, retrained from scratch6.49
Elastic Net6.67
Ours without graph7.85
AltumAge (published)4.88
AltumAge, fine-tuned5.14

For this new split strategy, it consists of 799 test samples from six studies, aged 0 to 91. Our model has the lowest MAE of all models without training overlap (4.35), ahead of the calibrated Horvath clock (5.24), the age-transformed Elastic Net (5.72), GraphAge (5.74) and DeepMAge (5.91), and it is also below the published AltumAge (4.88), whose training data likely include these samples. In addition, as shown in the table below, in the GSE19711 that includes 536 adults aged 49 to 91. Therefore, our model is robustness to study level shifts.

As for the graph-specific gain, according to the table below, our model is more accurate than our no-graph variant (4.35 vs 7.85 years) and in five of the six test studies. The largest difference is on E-GEOD-67444, where a third of the beta values are exactly zero (1.65 vs 29.34), consistent with message passing from correlated neighbors compensating for missing values. Without this study, our model is still more accurate (4.61 vs 5.78).

StudySamplesOursOurs without graph
E-GEOD-54399481.291.41
E-GEOD-65638168.6911.61
E-GEOD-67444701.6529.34
E-GEOD-83334303.156.29
GSE197115364.956.29
GSE37008994.184.05
Fold 2 overall7994.357.85

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.